Fast Statistical Models for Python

This repository contains the PyCon HK 2026 deck (first given at PyCon US 2026) and reproducibility artifacts for Breaking the Speed Limit: Python 3.14, Numba, and JAX in Statistical Computing.

The original PyCon US 2026 slides and poster are kept on the pyconus branch.

Public pages

The GitHub Pages version is the easiest way to view the talk materials:

Material	Open online	Source
Slides	View slides	slides/index.html
Poster	View poster	poster/index.html
Slides print mode	Open print/PDF view	slides/README.md

Core thesis:

• simulation is how statisticians write tests;
• speed only counts after the statistic is preserved;
• Python remains the scientific interface;
• Numba, JAX, Python 3.14, threads, processes, and A100 solve different bottlenecks.

Local preview

Use the project conda environment by default:

conda activate py312
python -m http.server 8000

Open the local pages:

• Slides: http://localhost:8000/slides/
• Poster: http://localhost:8000/poster/
• Slides print/PDF mode: http://localhost:8000/slides/index.html?print-pdf

The canonical deck is slides/index.html. Current structure is 31 slides total: 24 main-path slides and 7 hidden backup slides.

Results layout

• MacBook/local trust tier: experiments/results/macbook_air_long/latest/
• Linux server CPU scale tier: experiments/results/linux_server_cpu/long_safe_20260503_190133/
• Historical A100 long-safe tier: experiments/results/linux_server_a100/long_safe_20260503_190133/
• A100 permutation break-even notes: experiments/results/linux_server_a100/permutation_break_even/
• Slide-ready figures: experiments/results/presentation_figures/

Current result status

• MacBook k-means and permutation correctness/calibration results are the trust tier used in the main talk.
• Linux server CPU and k-means A100 results are scale evidence used in the main talk.
• The old A100 permutation matched slice is historical negative evidence: n=5,000, p=50,000, batch_R=512, before streamed reduction and the broader shape sweep.
• The current A100 permutation break-even narrative uses a100_streamed_reduction, larger batch_R, and a measured shape sweep. Speedup is scoped to the matched CPU matrix baseline divided by A100 streamed full end-to-end. Compile is excluded, transfer is included, and kernel-only rows are not used for CPU/A100 decisions.
• A100 permutation correctness wording is intentionally scoped: new accepted rows use pass_exact or pass_gpu_tolerance, while ambiguous rows use manual_check; historical rows marked check appear only in older follow-up notes.

Regenerate figures

MacBook figures:

python -m experiments.visualization.plot_macbook_air_evidence \
  --results-dir experiments/results/macbook_air_long/latest

Server CPU and k-means/A100 presentation figures:

python -m experiments.visualization.plot_server_talk_evidence

A100 permutation break-even figures, when the experiment CSVs are available:

python -m experiments.server.a100_permutation_break_even plot \
  --out-dir experiments/results/linux_server_a100/permutation_break_even \
  --presentation-dir experiments/results/presentation_figures

Deck QA

Review screenshots are local generated artifacts under slides/review/screenshots/. The QA pass checks browser mode, video poster fallbacks, broken media URLs, clipping, overlap, console errors, and A100 pending/experimental wording.